Del Prete, Roberto (2024) Intelligence Onboard Satellites: Deep Learning techniques for effective onboard data processing. [Tesi di dottorato]

Documento PDF DelPrete_Roberto_37.pdf Visibile a [TBR] Amministratori dell'archivio Download (202MB) | Richiedi una copia

Abstract

The timely and efficient onboard processing of vast quantities of data is crucial for the effective exploitation of current and future space sensing capabilities. Space missions generate enormous quantity of data which require rapid and tailored analysis to support real-time decision-making and enhance operational efficiency. In very recent times, the deployment of Artificial Intelligence (AI) onboard space missions has attracted considerable interest from the scientific community due to its substantial benefits, including reduced latency, minimized bandwidth, and improved system autonomy. By processing data directly on-the-fly, AI can significantly curtail the need for extensive data transmission to ground stations, thereby optimizing bandwidth utilization, expediting response times, and enhancing overall mission autonomy. This technology holds transformative potential for reshaping emergency response mechanisms and advancing space exploration, ultimately delivering significant benefits to humankind. Nonetheless, significant challenges persist in the implementation of machine learning for onboard space systems. A primary challenge lies in the considerable cost associated with data acquisition from space environments, which hinders the adoption of data-driven machine learning algorithms. Additionally, the unique physical and operational complexities inherent to the space environment—such as the impact of solar radiation on sensor performance, stringent constraints on spacecraft size, weight, and power, as well as restricted communication bandwidth for telemetry and data transfer—present challenges that are largely absent in terrestrial applications. Moreover, the domain gap problem, resulting from the inherent difficulty in generating datasets that comprehensively represent operational space scenarios, continues to impose substantial limitations on the effective automation of machine learning models for space applications. To address these challenges, this thesis explores the application of deep learning methodologies for onboard satellite data processing across three critical domains: wildfire response, autonomous spacecraft navigation, and maritime situational awareness. Each of these domains presents distinct challenges, such as the need for rapid detection and localization of wildfires from sparse data, the precise navigation of spacecraft in dynamic and unpredictable environments, and the monitoring of vast maritime regions for potential threats. Addressing these challenges requires the development of innovative AI-driven solutions that can enhance the efficiency, robustness, and autonomy of satellite operations, ultimately pushing the boundaries of what can be achieved with current space technology. By integrating AI capabilities directly into spaceborne systems, this research aims to substantiate the advantages of onboard intelligence, and the findings presented herein contribute to the advancement of AI-driven space sensing, providing a foundational framework for more autonomous and efficient satellite operations. This work not only underscores the technical feasibility of onboard AI but also highlights its transformative potential for future space missions, capable of independent operation in complex and evolving environments.

Downloads

Actions (login required)

Modifica documento